Add research report on the joint g-computation refactor

[smjenness]

A standalone ~2,800-word writeup documenting the joint g-computation refactor that the package shipped via PRs #66 through #77. Lives at inst/validation/method_refactor_report.md, parallel to the machine-generated comparison report method_comparison.md.

Structured as a research report:

• Introduction — ARTnet's role in EpiModelHIV-p; the marginal-vs-joint problem; the ARTnetPredict motivation for fixing the within-ARTnet baseline before applying ML-based forward projection
• Methods — the three new arguments (method, duration.method, target_pop); per-layer joint GLMs (Poisson, binomial, Gaussian, log-linear) with AIC-based interaction selection; g-computation aggregation in build_netstats; the deliberate choice that durations target mean age of extant ties at cross-section rather than mean full partnership duration (Steve Goodreau's framing in Length-biased sampling and 5-partnership truncation bias in formation target stats #72); validation infrastructure
• Results — headline finding: 229/363 cells (63%) shift > 5% across four scenarios; worst shifts on matched-and-old dissolution durations (−47%), older-age inst nodematch (−51%), high-deg.main casual nodefactor (+40%); concrete decomposition of the −15% Atlanta main-edges shift to the ARTnet vs Atlanta race composition gap; coefficient strengthening on deg.casl (−0.24 → −0.55), hiv2 (+0.09 → +0.25); end-to-end ERGM convergence verified
• Discussion — implications for EpiModelHIV-p (Atlanta-specific models over-target main edges by ~15% under legacy); three explicit limitations (geometric TERGM dissolution can't honor Weibull k ≠ 1, length-bias and 5-truncation not yet addressed in Length-biased sampling and 5-partnership truncation bias in formation target stats #72, joint_lm uses ongoing-only); ARTnetPredict's three unblocked next steps; outline for a methods paper
• References + Reproducibility snippet

Numbers cited spot-checked against the committed inst/validation/method_comparison.md for consistency. No code changes; documentation-only PR.

Test plan

• Word count in target range (2,811 / target 2,000–3,000)
• Numbers cited match inst/validation/method_comparison.md
• Sections cover the four requested areas (intro, methods, results, discussion)
• No code touched — pure docs PR

After this lands, the inst/validation/ directory holds:

• validate_backward_compat.R + snapshots — backward-compat regression infrastructure
• method_comparison.R + method_comparison.md — automated cross-method comparison tables
• method_refactor_report.md — this writeup
• epimodelhiv_template_ref/ — pinned downstream consumer scripts
• netstats_contract.md — field-level public contract
• README.md — workflow guide

Together that covers regression testing, cross-method numerics, and now the methodological narrative.